A New Tool for Automated CME Detection and

A New Tool for Automated CME Detection and Tracking with Deep Learning Pengyu Wang 1, Yan Zhang 1, Li Feng 2, Hanqing Yuan 1, and Jiahui Shan 2 1. Department of Computer Science and Technology, Nanjing University, China 2. Purple Mountain Observatory, Chinese Academy of Sciences, China lfeng@pmo. ac. cn

CMEs: major cause of geomagnetic storms • • … Disruption of electrical systems Communications Navigation systems Satellite hardware damage Online CME Catalogs Training Machine Learning Methods Region Tracking

Pipeline Online CME Catalogs Training Machine Learning Methods Region Detection Data Preprocessing Tracking

Pipeline Online CME Catalogs Training Machine Learning Methods Region Detection CNN Feature Extraction Tracking

Pipeline Online CME Catalogs Training Machine Learning Methods Region Detection Unsupervised CME Region Detection Tracking

Pipeline Online CME Catalogs Training Machine Learning Methods Region Detection Tracking CME in time series and refinement

Data Preprocessing LASCO C 2 running-difference images (from 2011 to 2017) resizing 256 x 256, denoising collecting CME Labels from online Catalogs for training binary classification CNN

CNN Feature Extraction convolutional neural networks have shown excellent performance on many computer vision tasks CNN helps to learn more robust image features Object detection Human understanding Autonomous driving

CNN Feature Extraction we train a Le. Net model (3 convolution layers, 2 downsampling layers and 2 fully-connected Layers) input 256 x 256 image, output the probability of a CME event For training, we obtain whether a frame is labeled with or without a CME from online Catalogs for data of a few months Le. Net Model

CME Region Detection Each image's CNN feature map contains high level semantic information of CME region Use PCA (Principal Component Analysis) to localize the CME region in an image set

PCA Co-localization PCA converts correlated variables into uncorrelated variables PCA transforming [Wei et al. 2017] The uncorrelated variable with the largest variance can be treated as an indicator of CME region

PCA Co-localization Results

Refinement graph cut method finely tune the CME region boundary consider CME region probability and neighbor information graph cut origin image CME region probability refinement results

Le. Net Comparison of the detection CACTUS graphcut SEEDS CACTUS

Tracking new CME moves outward in at least two running-difference images [Olmedo et al. 2008] robust to noise feature Refinement + Tracking new CME(in color Blue)

Tracking convert to polar coordinate system compute starting/ending position angles and maximal height for each angle polar coordinate

Tracking consider next 2 running-difference images Frame i-1 Frame i+1

Tracking consider next 2 running-difference images compute angular width overlap: intersection set /union set > 0. 1 noise Frame i-1 Frame i+1

Tracking consider next 2 running-difference images compute angular width overlap: intersection set /union set > 0. 1 new CME Frame i-1 Frame i+1

Summary and Outlook • We have developed a new tool for automated CME detection and tracking using the deep learning technique. CAMEL • An online catalog is being built with the key parameters of CMEs. 2011. 2. 15 -2: 34 2017. 9. 06 -11: 12 2011. 12. 03 -11: 09 2011. 12. 29 -16: 35